Assessment of multi-carbon stocks and influencing factors in terrestrial ecosystems based on ensemble learning algorithm

Understanding the spatiotemporal dynamics and drivers of terrestrial ecosystem carbon storage is essential for sustaining the global carbon balance and mitigating climate change. However, limited carbon density data and the poor generalization of inversion models have confined regional assessments to a single carbon pool, obscuring their driving mechanisms. To address these gaps, we developed an integrated framework combining Extreme Gradient Boosting (XGBoost), SHapley Additive exPlanations (SHAP), and Structural Equation Modeling (SEM), and applied it to Shandong Province, China. Combined with Theil–Sen slope estimation, Mann–Kendall trend tests, the coefficient of variation, and the Hurst index, we quantified the spatiotemporal evolution of aboveground (AGC), belowground (BGC), soil organic (SOC), and total (TOC) carbon stocks from 2000 to 2020 and identified key environmental drivers. Results showed that the models performed well, with SOC achieving the highest accuracy (R2 = 0.84), followed by AGC (R2 = 0.66) and BGC (R2 = 0.63). TOC increased from 887.77 to 910.98 TgC, with SOC contributing 40–50% and AGC showing the fastest growth. Carbon stocks displayed a “central-east high, peripheral low” pattern, shifting toward ecological barrier zones. Cropland loss (1.15 × 104 km2) reduced TOC by 88.40 TgC, while forest and built-up land expansion increased TOC by 112.88 TgC. Vegetation factors primarily affected AGC, whereas BGC and SOC were more responsive to topography and climate; human activities exerted negative effects on all pools. These findings underscore the importance of a multi-pool perspective for ecosystem carbon dynamics and provide methodological and strategic guidance for regional carbon sink assessment and low-carbon development.